JP4976641B2 - Method and apparatus for recommending target items based on third party stereotype preferences - Google Patents

Abstract

A method and apparatus are disclosed for recommending items of interest to a user, such as television program recommendations, before a viewing history or purchase history of the user is available. A third party viewing or purchase history is processed to generate stereotype profiles that reflect the typical patterns of items selected by representative viewers. A user can select the most relevant stereotype(s) from the generated stereotype profiles and thereby initialize his or her profile with the items that are closest to his or her own interests. A clustering routine is disclosed to partition the third party viewing or purchase history (the data set) into clusters, such that points (e.g., television programs) in one cluster are closer to the mean of that cluster than any other cluster. A mean computation routine is also disclosed to compute the symbolic mean of a cluster.

Description

  The present invention relates to a method and apparatus for recommending target items, such as television programs, and more particularly to a method for recommending programs and other target items before a user's purchase or viewing history is available.

  As the number of channels available to TV viewers increases in addition to the variety of program content available on such channels, it becomes increasingly difficult for TV viewers to distinguish between target TV programs. Yes. Electronic Program Guide (EPG) identifies available TV programs by, for example, title, time, date, and channel, and allows you to search or sort available TV programs according to personal preferences This makes it easy to specify the target program.

A number of recommendation tools have been proposed or proposed that recommend television programs and other target items. For example, the television program recommendation tool applies a viewer preference to an EPG and acquires a recommended program group that a specific viewer may be interested in. In general, television program recommendation tools obtain viewer preferences using explicit or implicit tools, or some combination of the above. The implicit television program recommendation tool generates a television program recommendation unknowingly based on information derived from the viewing history of the viewer. Explicit TV program recommendation tools, on the other hand, explicitly ask viewers about their preferences about program attributes, such as title, genre, actor, channel and date and time, to derive a viewer profile and generate recommendations .
US Patent Application Publication No. 09/819286 U.S. Patent Application Publication No. 09/466406 US Patent Application Publication No. 09/498271 U.S. Patent Application No. 09/627139 Stanfill and Waltz, “Toward Memory-Based Reasoning”, Communications of the ACM, 1986, 29:12, pages 1213-1228 Cost and Salzberg, “A Weighted Nearest Neighbor Algorithm For Learning With Symbolic Features”, Machine Learning, Kluwer Publishers (Boston, MA), 1993, 10, pp. 57-58 J. Kittler et al., “Combing Classifiers”, Proc. Of the 13th Int'l Conf. On Pattern Recognition, Austria, Vienna, 1996, II, pages 897-901

  Currently available recommendation tools help users identify target items, but are subject to a number of limitations, which, when overcome, can greatly improve the convenience and performance of such recommendation tools . For example, in a thorough implementation of an explicit recommendation tool, initialization is very verbose and each new user responds to a very detailed survey that identifies the new user's preferences at a coarse-grain level. I need that. While implicit television program recommendation tools derive profiles by observing viewing behaviors without knowing it, it takes a long time to be accurate. Furthermore, such an implicit television program recommendation tool requires at least a minimum amount of viewing history to begin making any recommendations. Thus, such an implicit television program recommendation tool cannot make any recommendations when the recommendation tool is first obtained.

  Accordingly, there is a need for a method and apparatus that can recommend items such as television programs without the knowledge before sufficient personalized viewing history is available. Furthermore, there is a need for a method and apparatus for generating program recommendations for specific users based on third party viewing behavior.

  Generally, a method and apparatus for recommending a target item, such as a recommended television program, to a user is disclosed. According to one aspect of the present invention, recommendations may be generated before a user's viewing history or purchase history is available, such as when the user first obtains the recommendation tool. First, a viewing history or purchase history from one or more third parties is used to recommend a target item for a specific user. Third party viewing or purchase history is obtained from a sample population with demographics representative of a larger population.

  The third party viewing or purchase history is processed to generate a stereotype profile that reflects the typical pattern of items selected by the typical viewer. As used herein, a stereotype profile is a cluster of items (data points) that are somewhat similar to each other. Thus, a particular cluster corresponds to a particular segment of selected items from third party viewing or purchase history that represents a particular pattern. Once the stereotype profile has been generated by the present invention, the user can select the most appropriate stereotype, thereby initializing the user's profile with the item closest to the user's own interests. . The stereotype profile is further tailored and developed by providing selection patterns and some feedback for each individual user's specific personal viewing behavior.

  A cluster in which a third party viewing or purchase history (data set) is divided into clusters and the points in one cluster (eg, a television program) are closer to the average value of that cluster than any other cluster Disclosed are routines. Further, an average value calculation routine for calculating the symbolic average value of the cluster is disclosed. A particular data point, such as a television program, is assigned to a cluster using the average value of each cluster based on the distance between the data point and each cluster. The clustering performance evaluation routine determines when a predefined criterion that stops cluster generation is met.

  FIG. 1 shows a television program recommendation tool 100 according to the present invention. As shown in FIG. 1, the exemplary television program recommendation tool 100 evaluates programs in the program database 200 described in conjunction with FIG. 2 to identify targeted programs for specific audiences. The recommended program group is displayed to the viewer using, for example, a well-known on-screen presentation method using a set top terminal / TV (not shown). While the present invention is illustrated herein in connection with television program recommendations, the present invention may be applied to any automatically generated recommendations based on evaluation of user behavior, such as viewing history or purchase history.

  According to one aspect of the present invention, the TV program recommendation tool 100 recommends a TV program before the user's viewing history 140 is available, such as when the user first acquires the TV program recommendation tool 100. Can be generated. As shown in FIG. 1, the television program recommendation tool 100 first recommends a target program to a specific user using a viewing history 130 from one or more third parties. In general, the third party viewing history 130 is based on the viewing habits of one or more sample populations having demographics representative of the larger population, such as age, income, gender and education.

  As shown in FIG. 1, the third-party viewing history 130 includes viewing / non-viewing program groups based on a specific population. The viewing program group is acquired by observing a program actually viewed by the specific population. The program group that is not viewed is acquired by, for example, randomly sampling programs in the program database 200. There is another example in which a program group that is not viewed is acquired by teaching of a patent application dated March 28, 2001 (see Patent Document 1) entitled “Adaptive Sampling Technique for Selecting Negative Examples for Artificial Intelligence Applications”.

  According to another aspect of the present invention, the television program recommendation tool 100 generates a stereotype profile that reflects typical patterns of television programs viewed by typical viewers. Process 130. As described further below, a stereotype profile is a cluster of television programs (data points) that are somewhat similar to each other. Thus, a particular cluster corresponds to a particular segment of a television program from the third party viewing history 130 that represents a particular pattern.

  Third party viewing history 130 is processed by the present invention to provide a cluster of programs representing a particular pattern. The user can then select the most appropriate stereotype, thereby initializing the user's profile with the program closest to the user's own interests. Stereotype profiles are further developed and tailored to specific individual viewing behaviors of each individual user in response to their recording patterns and feedback to the program. In one embodiment, in determining the program score, a program from a user's own viewing history 140 may be given a higher weight than a program from a third party viewing history.

  The television program recommendation tool 100 is a personal computer having a processing unit 115 such as a central processing unit (CPU) and a storage unit 120 such as a RAM (read / write storage unit) and / or a ROM (read only storage unit). It can be implemented as any computing device, such as a computer or workstation. The television program recommendation tool 100 may further be implemented as an application specific integrated circuit (ASIC), for example in a set top terminal (not shown) or display device. Further, the TV program recommendation tool 100 is any available TV program recommendation tool, such as the system commercially available from Tivo ™, Tivo, Inc. (Sunnyvale, Calif.). A patent application dated December 17, 1999 (see Patent Document 2) entitled "Method and Apparatus for Recommending Television Programming Using Decision Trees", 2000, entitled "Bayesian TV Show Recommender" The television program described in the patent application dated February 4, 2000 (see Patent Document 3) and the patent application dated July 27, 2000 AD (see Patent Document 4) entitled “Three-Way Media Recommendation Method and System” As a recommended tool, or any combination thereof, it may be implemented as modified herein to carry out the features and functions of the present invention. .

  As shown in FIG. 1 and further described below in conjunction with FIGS. 2-8, the television program recommendation tool 100 includes a program database 200, stereotype profile processing 300, clustering routine 400, average value calculation routine 500, distance calculation routine. 600 and a cluster performance evaluation routine 800. In general, program database 200 may be implemented as a well-known electronic program guide and records information about each program available at a particular time interval. Stereotype profile processing 300 (i) processes third party viewing history 130 to generate a stereotype profile that reflects typical patterns of television programs viewed by typical viewers; (ii) ) Allows the user to select the most appropriate stereotype; and (iii) generates recommendations based on the selected stereotype.

  Clustering routine 400 clusters third party viewing history 130 (data set) so that the points (television programs) in one cluster are closer to the average (center of mass) of the cluster than any other cluster. Called by the stereotype profiling process 300 to divide into The clustering routine 400 calls an average value calculation routine 500 to calculate the symbolic average value of the cluster. The distance calculation routine 600 is called by the clustering routine 400 to evaluate how close a television program is to each cluster based on the distance between the particular television program and the average value of the particular cluster. Finally, the clustering routine 400 calls the clustering performance evaluation routine 800 to determine when the criteria for stopping cluster generation has been met.

  FIG. 2 is a sample table from the program database (EPG) 200 of FIG. As described above, the program database 200 records information about each program available at a specific time interval. As shown in FIG. 2, the program database 200 has a plurality of records, such as records 205-220, each associated with a particular program. For each program, program database 200 indicates the date and time and channel associated with the program in fields 240 and 245, respectively. In addition, the title, genre and actor of each program are specified in fields 250, 255 and 270. Other well-known features (not shown) may also be included in the program database 200, such as program length and description.

  FIG. 3 is a flow diagram illustrating an exemplary embodiment of a stereotype profile process 300 incorporating features of the present invention. As described above, the stereotype profile processing 300 (i) sets the third party viewing history 130 to generate a stereotype profile that reflects the typical pattern of a television program viewed by a typical viewer. Processing; (ii) allows the user to select the most appropriate stereotype and thereby initialize the user's profile; and (iii) generate recommendations based on the selected stereotype . The processing of the third party viewing history 130 can be performed offline, for example, in a factory, and the TV program recommendation tool 100 can be provided for a user who has installed the generated stereotype profile for selection by the user. Special mention.

  Accordingly, as depicted in FIG. 3, stereotype profile processing 300 first collects third party viewing history 130 during step 310. The stereotype profile process 300 then executes a clustering routine 400 during step 320 to generate a cluster of programs corresponding to the stereotype profile, as described below in conjunction with FIG. As described further below, the exemplary clustering routine 400 uses an unmanaged data clustering algorithm for the viewing history data set 130, such as a “k-means” clustering routine. obtain. As described above, the clustering routine 400 allows the third party viewing history 130 (data- Set) is divided into clusters.

  Stereotype profile processing 300 further assigns a label characterizing each of the one or more stereotype profiles to each cluster during step 330. In one exemplary embodiment, the average value of the cluster becomes a typical television program for the entire cluster, and the features of the average value program can be used to label the cluster. For example, the television program recommendation tool 100 may be configured such that the genre is the dominant, i.e., characterizing, feature for each cluster.

  The labeled stereotype profile is presented to each user during step 340 for selection of the stereotype profile that most closely matches the user's interest. The programs that make up each selected cluster can be thought of as a “typical viewing history” of that stereotype and can be used to build a stereotype profile for each cluster. Accordingly, during step 350, a viewing history having programs from the selected stereotype profile is generated for the user. Finally, the viewing history generated in the above step is applied to the program recommendation tool during step 360 to obtain program recommendations. The program recommendation tool may be implemented as any conventional program recommendation tool modified herein, as referenced above, as will be apparent to those skilled in the art. Control of the program ends during step 370.

  FIG. 4 is a flow diagram illustrating an exemplary embodiment of a clustering routine 400 incorporating features of the present invention. As described above, the clustering routine 400 is invoked during step 320 by the stereotype profile processing 300 so that points (television programs) in one cluster are the average (center of mass) of that cluster over any other cluster. The third party viewing history 130 (data set) is divided into clusters so as to be close to. In general, clustering routines focus on unmanaged tasks that find example groupings in a sample data set. The present invention divides the data set into k clusters using a k-means clustering algorithm. As described below, the two main parameters for the clustering routine 400 include: (i) a distance metric to find the nearest cluster, described below in conjunction with FIG. 6; and (ii) the number of clusters to generate, k; is there.

  The example clustering routine 400 uses the dynamic value of k, provided that further clustering of the example data does not result in any improvement in classification accuracy, reaching a stable value k. In addition, the cluster size is increased to the point where an empty cluster is recorded. Thus, clustering stops when the natural level of the cluster is reached.

  As shown in FIG. 4, the clustering routine 400 first determines k clusters during step 410. The example clustering routine 400 begins by selecting the minimum number of clusters, eg, two. For this fixed number, the clustering routine 400 processes the entire viewing history data 130 and is stable after several iterations (ie, moving from one cluster to another even if the algorithm is repeated once more). Two clusters that can be considered) are reached. The current k clusters are initialized during step 420 by one or more programs.

  In one exemplary embodiment, the cluster is initialized with a number of seed programs selected from third party viewing history 130 during step 420. Programs that initialize the cluster may be selected randomly or sequentially. In a sequential embodiment, a cluster may be initialized with a program that begins with the first program in the viewing history 130 or a program that begins at any point in the viewing history 130. In yet another variation, the number of programs that initialize each cluster may be further varied. Finally, the cluster may be initialized with one or more “virtual” programs having feature values randomly selected from programs in the third party viewing history 130.

  Thereafter, the clustering routine 400 initializes an average value calculation routine 500, described below in conjunction with FIG. 5, to calculate the current average value of each cluster during step 430. The clustering routine 400 further executes a distance calculation routine 600, described below in conjunction with FIG. 6, to determine the distance to each cluster of each program in the third party viewing history 130 during step 440. Each program in viewing history 130 is further assigned to the nearest cluster during step 460.

  During step 470, a test is performed to determine if there are programs that have moved from one cluster to another. If it is determined during step 470 that the program has moved from one cluster to another, program control returns to step 430 and continues in the manner described above until a stable group of clusters is identified. However, if it is determined during step 470 that no program has moved from one cluster to another, program control proceeds to step 480.

  During step 480, a further test is performed to determine whether a particular performance criterion has been met or an empty cluster has been identified (collectively referred to as a “stop criterion”). If the stop criteria is not met during step 480, the value of k is increased during step 485 and program control returns to step 420 and continues in the manner described above. However, if it is determined during step 480 that the stop criteria have been met, program control ends. The evaluation of the stop criterion will be further described below with reference to FIG.

  The exemplary clustering routine 400 places programs into one cluster, thus creating what is called a crisp cluster. Yet another variation utilizes fuzzy clustering, thereby allowing a particular example (television program) to partly belong to many clusters. In a fuzzy clustering method, television programs are assigned a weight, which represents how close the television program is to the cluster average. The weighting may depend on the inverse square of the distance of the television program from the cluster average. The sum of all cluster weights associated with a single television program should be 100%.

Cluster Symbolic Average Calculation FIG. 5 is a flow diagram illustrating an exemplary embodiment of an average calculation routine 500 incorporating features of the present invention. As described above, the average value calculation routine 500 is called by the clustering routine 400 to calculate the symbolic average value of the cluster. For numerical data, the average value is the value that minimizes the variance value. By extending this notion to symbolic data, the mean value of the cluster can be specified by finding the value of x _μ that minimizes the intracluster variance (and hence the radius or region of the cluster)

で、Jは同様のクラス（視聴又は非視聴）からのテレビ番組のクラスタで、ｘ_iは番組iについての記号的特徴値で、x_μはVar(J)を最小化するような、Jにおけるテレビ番組の１つからの特徴値である。

Where J is a cluster of television programs from similar classes (viewing or non-viewing), x _i is a symbolic feature value for program i, and x _μ is such that it minimizes Var (J). A feature value from one of the television programs.

Thus, as shown in FIG. 5, the average calculation routine 500 first identifies the program currently in the particular cluster J during step 510. For the current symbolic attribute of interest, in step 520, the variance value of the cluster is calculated for each possible symbolic value x _μ using equation (1). The symbolic value x _μ that minimizes the variance value is selected as an average value during step 530.

  A test is performed to determine if there are more symbolic attributes to consider during step 540. If it is determined in step 540 that there are more symbolic attributes to consider, program control returns to step 520 and continues in the manner described above. However, if it is determined in step 540 that there are no more symbolic attributes to consider, program control returns to the clustering routine 400.

For calculation, each symbolic feature value in J is tried as x _μ , and the symbolic value that minimizes the variance value is the average value of the target symbolic attribute in cluster J. Two types of average values can be calculated: feature-based average and program-based average.

Feature-Based Symbolic Average Value The exemplary average value calculation routine 500 described herein is feature-based, and the resulting cluster average value has feature values taken from the example (program) in cluster J, which is symbolic. This is because the average value of the target attribute must be one of its possible values. However, it is noted that the cluster average value can be a “virtual” television program. The feature value of this virtual program is taken from the channel value taken from one of the examples (eg “EBC”) and from another example (eg “BBC World News” which is not actually broadcast on “EBC”). May have a subject value. Thus, any feature value that represents the minimum variance value is selected to represent the average value of that feature. The average calculation routine 500 is repeated during step 540 until it is determined that all features (ie, symbolic attributes) have been considered for all feature locations. The resulting virtual program so obtained is used to represent the average value of the cluster.

Program-Based Symbolic Average In another variation, in equation (1) for the variance value, x _i can be the television program i itself, and similarly x _μ is the variance value for the programs in cluster J. This is a program in cluster J that minimizes. In this case, the metric is appropriate for minimizing the distance between programs rather than between individual specific values. Furthermore, in this case, the resulting average value is not a virtual program, but just a program selected from the J group. Any program found in cluster J that minimizes the variance value for all programs in cluster J can be used to represent the average value of the cluster.

Symbolic Average Using Multiple Programs The exemplary average calculation routine 500 described above uses a single specific value for each possible feature (whether feature-based or program-based embodiments). Characterize the cluster average. However, it has been found that relying on only one feature value for each feature during mean value calculation often leads to inadequate clustering because the mean value is no longer at the typical cluster center for the cluster. . That is, it may not be desirable to represent a cluster with only one program, but rather multiple programs representing an average value or multiple average values may be utilized to represent a cluster. Thus, in another variation, a cluster may be represented by multiple average values or multiple feature values for each possible feature. Thus, N features (for feature-based symbolic averages) or N programs (for program-based symbolic averages) that minimize the variance value are selected during step 530, where N is The number of programs used to represent the average value.

Calculation of Distance Between Programs and Clusters As described above, the distance calculation routine 600 is called by the clustering routine 400 to determine how close a TV program is to each cluster and the average value of a specific TV program and a specific cluster. Evaluate based on the distance between. The calculated distance metric quantifies the differences between the various examples in the sample data set to determine the area of the cluster. In order to be able to cluster user profiles, it is necessary to calculate the distance between any two television programs in the viewing history. In general, television programs that are close to each other tend to be in one cluster. There are a number of relatively simple techniques for calculating the distance between numerical vectors, such as the Euclidean distance, the Manhattan distance, and the Mahalanobis distance. Existing distance calculation methods, however, cannot be used in the case of television program vectors, since television programs mainly have symbolic feature values. For example, the story of “Devil” broadcast on “EBC” on March 22, 2001 at 8 pm and the story of “Simons” broadcast on “FEX” on March 25, 2001 at 8 pm Two such television programs may be represented using the following feature vectors:
Title: Devil Title: Simons Channel: EBC Channel: FEX
Broadcast date: March 22, 2001 Broadcast date: March 25, 2001 Broadcast time: 20:00 Broadcast time: 20:00
Obviously, a known numerical distance metric cannot be used to calculate the distance between the feature value “EBC” and the feature value “FEX”. Value difference metric (VDM) is an existing technique for measuring the distance between feature values in a symbolic feature value region. The VDM approach takes into account the classification similarity of all instances for each possible value of each feature. By using this method, a matrix that identifies the distance between all feature values is statistically derived based on examples in the training group. Some VDM methods for calculating the distance between symbolic feature values are described in more detail (see, for example, Non-Patent Document 1).
The present invention utilizes the VDM technique or variations thereof to calculate the distance between feature values between two television programs or other target items. The original VDM uses a weighting term in calculating the distance between two feature values, thus making the distance metric asymmetric. The modified VDM (MVDM) makes the distance matrix symmetric by excluding the weighting term. Some MVDM methods for calculating the distance between symbolic feature values are described in more detail (for example, see Non-Patent Document 2).
According to MVDM, the distance δ between the two values of V1 and V2 is

によって示される。本発明の番組推奨環境においては、MVDM式(3)は、特に「視聴」及び「非視聴」のクラスを扱うよう変換される。

Indicated by. In the program recommendation environment of the present invention, the MVDM expression (3) is converted to handle classes of “viewing” and “non-viewing”.

式(4)においては、V1及びV2は対象とする特徴について考えられる2つの値である。上記の例について続けると、特徴「チャンネル」について、第１値V1は「EBC」に等しくて、第２値V2は「FEX」に等しい。該値の間の距離は当該例が分類される全てのクラスの和である。本発明の例示的番組推奨ツール実施例について適切なクラスは「視聴」及び「非視聴」である。C1iはV1（「EBC」）がクラスi(iはクラス視聴を意味する1に等しい)に分類された回数で、C1(「C1_total」)はV1がデータ・セット中に発生した合計回数である。値「r」は定数で、通常１に設定される。

In Equation (4), V1 and V2 are two possible values for the target feature. Continuing with the above example, for the feature “channel”, the first value V1 is equal to “EBC” and the second value V2 is equal to “FEX”. The distance between the values is the sum of all classes into which the example is classified. Suitable classes for the exemplary program recommendation tool embodiment of the present invention are “viewing” and “non-viewing”. C1i is the number of times V1 ("EBC") is classified as class i (i is equal to 1 meaning class viewing) and C1 ("C1_total") is the total number of times V1 occurred in the data set . The value “r” is a constant and is usually set to 1.

  If the metric defined by equation (4) occurs with similar relative frequency for all classifications, the value is specified as similar. The term C1i / C1 represents the possibility that the residual median is classified as i, assuming that the feature of interest has the value V1. Thus, two values are similar if they indicate similarities for all possible classifications. Equation (4) calculates the overall similarity between two values by finding the sum of the possible differences of the values in all classifications. The distance between two television programs is the sum of the distances between corresponding feature values of the two television program vectors.

  FIG. 7A is a portion of a distance table for feature values related to the feature “channel”. FIG. 7A programs the number of occurrences of each channel feature for each class. The values depicted in FIG. 7A are taken from an exemplary third-party viewing history 130.

  FIG. 7B displays the distance between each feature value pair calculated from the exemplary statistics depicted in FIG. 7A using MVDM equation (4). Intuitively speaking, “EBC” and “ABS” mostly occur in the class “viewing” and do not occur in the class “non-viewing” (“ABS” has a slightly non-viewing component. So, they should be “close” to each other. FIG. 7B enhances intuition by the fact that the distance between “EBC” and “ABS” is a small (non-zero) value. “ASPN”, on the other hand, occurs mostly in the class “non-viewing”, so this data set should be “far” from both “EBC” and “ABS”. FIG. 7B is programmed such that the distance between “EBC” and “ASPN” is 1.895, for a maximum possible distance of 2.0. Similarly, the distance between “ABS” and “ASPN” is high at a value of 1.828.

  Accordingly, as shown in FIG. 6, the distance calculation routine 600 first identifies the program in the third party viewing history 130 in step 610. For the current program of interest, the distance calculation routine 600 uses equation (4) to calculate the characteristics of each symbolic feature value during step 620 up to the feature corresponding to each cluster average value (determined by the average value calculation routine 500). Calculate the distance.

  The distance between the current program and the cluster average is calculated during step 630 by summing the distances between corresponding feature values. During step 640, a test is performed to determine whether there are more programs of interest in the third party viewing history 130. If it is determined in step 640 that there are more programs of interest in the third party viewing history 130, the next program is identified in step 650 and program control proceeds to step 620 and continues in the manner described above. .

  However, if it is determined in step 640 that there are no more programs in the third party viewing history 130, program control returns to the clustering routine 400.

  Like the “symbolic mean derived from multiple programs” above, the mean value of the cluster is the feature value for each possible feature (regardless of feature-based or program-based embodiment). Can be characterized using numbers. Results resulting from multiple averages are further pooled to reach consensus decisions through voting by a variation of the distance calculation routine 600. For example, now in step 620, the distance between a particular feature value of the program and each of the feature values corresponding to various average values is calculated. The minimum distance results are pooled and used for voting, for example by voting majority or experts to reach consensus decisions. Some have described such a technique in more detail (see, for example, Non-Patent Document 3).

Stop Criteria As described above, the clustering routine 400 calls the clustering performance evaluation routine 800, as shown in FIG. 8, to determine when the criteria for stopping cluster generation have been met. The example clustering routine 400 uses the dynamic value of k, provided that further clustering of the example data does not result in any improvement in classification accuracy and reaches a stable value k. Furthermore, the cluster size can be increased to the point where empty clusters are recorded. Thus, clustering stops when the natural level of the cluster is reached.

  The example clustering performance evaluation routine 800 tests the classification accuracy of the clustering routine 400 using a subset of programs (test data set) from the third party viewing history 130. For each program in the test set, the clustering performance evaluation routine 800 determines the cluster closest to the program (which cluster average is the closest) and compares the class labels for the cluster and the target program. The percentage of consistent class labels is interpreted as the accuracy of the clustering routine 400.

  Thus, as shown in FIG. 8, the clustering performance evaluation routine 800 first collects a subset of programs from the third party viewing history 130 to act as a test data set in step 810. . A class label is then assigned to each cluster during step 820 based on the percentage of viewing and non-viewing programs in the cluster. For example, if most of the programs in a cluster are being viewed, the cluster is assigned a “view” label.

  The cluster closest to each program in the test set is identified during step 830 and the class label for the assigned cluster is compared as to whether the program was actually viewed. In embodiments that use multiple programs to represent the average value of the cluster, an average distance (to each program) or a voting technique may be used. The consistency class level percentage is determined in step 840 before program control returns to the clustering routine 400. The clustering routine 400 ends when the classification accuracy reaches a predefined threshold.

  It should be noted that the embodiments and variations shown and described herein are merely illustrative of the principles of the invention and that various modifications may be made by those skilled in the art without departing from the scope and spirit of the invention. To do.

It is a schematic block diagram of the television program recommendation tool by this invention. 2 is a sample table from the exemplary program database of FIG. 2 is a flow diagram illustrating the stereotype profiling process of FIG. 1 embodying the principles of the present invention. 2 is a flow diagram illustrating the clustering routine of FIG. 1 implementing the principles of the present invention. 2 is a flowchart illustrating the average value calculation routine of FIG. 1 implementing the principles of the present invention. 2 is a flow diagram illustrating the distance calculation routine of FIG. 1 implementing the principles of the present invention. 6 is a sample table from an exemplary channel feature value generation table showing the number of occurrences of feature values for each channel for each class. 7B is a sample table from an exemplary feature value versus distance table showing the distance between each feature value pair calculated from the exemplary statistics depicted in FIG. 7A. 2 is a flow diagram illustrating the clustering performance evaluation routine of FIG. 1 implementing the principles of the present invention.

Claims (9)

A method for generating a user profile indicative of user preferences in a system having a processing unit comprising:
The processing device obtaining a third party selection history representing a television program selected by at least one third party;
The processing device comprising dividing the third party selection history into clusters of television programs ; and the processing device receiving at least one selection of the clusters from the user ;
As before SL user profile, a method of pre-Symbol viewing history with a television program from at least one selected cluster, wherein Rukoto generated by the processing device towards the user. The method of claim 1, further comprising:
The processing device recommending a television program based on the user profile;
A method characterized by comprising: The method of claim 1, further comprising:
The processing device assigning a label to each of the clusters;
A method characterized by comprising:   4. The method of claim 3, wherein the user selects the at least one cluster based on the assigned label. The method of claim 1, wherein the dividing step further comprises:
The processor uses a k-means clustering routine;
A method characterized by comprising: A system that generates a user profile that shows user preferences:
A storage device storing a computer readable code; and a processing device operatively coupled to the storage device;
The processing device has:
Obtaining a third party selection history indicating a television program selected by at least one third party;
Dividing the third party selection history into clusters of television programs ; and receiving at least one selection of the clusters from the user;
In addition,
System Examples user profile, the prior SL viewing history with a television program from at least one selected cluster towards the user, and wherein the Rukoto generated by said processing device. 7. The system of claim 6 , wherein the processing device is:
Recommending television programs based on the selected clusters;
A system characterized by being configured to perform: A device that generates a user profile indicating user preferences:
A computer-readable recording medium;
The medium has computer-readable code means on the medium, and the computer-readable code means is:
A function of obtaining a third party selection history indicating a television program selected by at least one third party;
A function of dividing the third party selection history into clusters of television programs ; and a function of receiving at least one selection of the clusters from the user;
In addition,
Wherein as the user profile, prior Symbol least one viewing history with a television program from the selected clusters toward the user, and wherein the Rukoto generated by readable code means by the computer. 9. The apparatus of claim 8 , further comprising code means readable by the computer:
The ability to recommend television programs based on the selected cluster;
A device characterized by comprising:

Images